AI Papers Reader

Personalized digests of latest AI research

View on GitHub

Open Deep Search: An Open-Source AI Democratizing Access to Advanced Search

San Francisco, CA – March 27, 2025 – In a move aimed at leveling the playing field in the rapidly evolving landscape of AI-powered search, a team of researchers has unveiled Open Deep Search (ODS), an open-source framework designed to bring state-of-the-art search capabilities to a wider audience. The project seeks to replicate the capabilities of closed-source search AI solutions.

ODS addresses the widening gap between proprietary search AI solutions like Perplexity’s Sonar Reasoning Pro and OpenAI’s GPT-4o Search Preview and their open-source counterparts. The core innovation lies in augmenting the reasoning abilities of Large Language Models (LLMs) with reasoning “agents” that can strategically employ web search tools to answer user queries.

Imagine you’re trying to find the answer to a complex question, like “What year was the band leader of the group who originally performed the song sampled in Kanye West’s song Power born?” Current open-source solutions might struggle, returning irrelevant snippets of text. ODS, on the other hand, uses a “Reasoning Agent” to break down the question: identify the sampled song, the original band, the band leader, and then their birth year. It uses a specialized “Open Search Tool” to find relevant information for each sub-query.

ODS comprises two main components:

  1. Open Search Tool: This tool is designed to improve upon existing open-source search tools by rephrasing queries when necessary, extracting relevant context from the top search snippets, and employing chunking and re-ranking to filter out irrelevant information. It can also handle website-specific APIs for major websites like Wikipedia and ArXiv.
  2. Open Reasoning Agent: This agent interprets the given task, assesses retrieved context, and uses appropriate tools, including the Open Search Tool, to formulate an answer. Two versions of the reasoning agent are provided, one based on the ReAct framework and the other on the CodeAct framework.

According to the researchers, ODS, when paired with open-source LLMs like DeepSeek-R1, can match or even surpass the performance of leading closed-source alternatives on benchmarks like SimpleQA and FRAMES. For example, on the FRAMES benchmark, ODS improves upon the recently released GPT-4o Search Preview by 9.7% in accuracy. In one example from the paper, ODS-v1 correctly identifies 112 inches as the correct wheelbase measurement of the Jensen Interceptor (1950) and makes an additional check with the Wolfram-Alpha API to convert the answer to 2,845mm. Conversely, Perplexity Sonar Reasoning Pro gets confused between 2,858mm and 2,845mm.

The research paper highlights the advantages of ODS in several concrete scenarios. For instance, ODS can accurately calculate the age difference in years between two events, correctly identifying the band leader of King Crimson when other solutions struggle, and discerning the correct date of a news event when conflicting information is present.

The open-source nature of ODS promotes transparency, innovation, and entrepreneurship in the field of search AI. The team hopes that this framework will enable a wider community of developers to contribute to the advancement of search technology.

The ODS code is available on Github at https://github.com/sentient-agi/OpenDeepSearch.